Prev: virtio-pci: disable msi at startup
Next: [PATCH 1/1] perf record: Don't call newt functions when not initialized
From: Linus Torvalds on 10 Jun 2010 12:00
On Thu, Jun 10, 2010 at 6:44 AM, Jens Axboe <jaxboe(a)fusionio.com> wrote:
>
> - A set of patches fixing the WB_SYNC_NONE writeback from Christoph. So
> �we should finally have both functional and working WB_SYNC_NONE from
> �umount context.

I _really_ think this is too late, considering how broken it has been.
We already reverted the WB_SYNC_NONE things exactly because it didn't
work, didn't we? I'm going to be off-line in two days, and this part
of the pull request really makes me nervous, if only simply because of
the history of it all (ie it's always been broken, why shouldn't it be
broken now?).

IOW, that's a lot of scary changes, that have historically not been
safe or sufficiently tested, and have caused problems for various
filesystems. Convince me why they should suddenly be ok to merge?

Linus
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo(a)vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/
From: Jens Axboe on 10 Jun 2010 12:30
On 2010-06-10 17:55, Linus Torvalds wrote:
> On Thu, Jun 10, 2010 at 6:44 AM, Jens Axboe <jaxboe(a)fusionio.com> wrote:
>>
>> - A set of patches fixing the WB_SYNC_NONE writeback from Christoph. So
>> we should finally have both functional and working WB_SYNC_NONE from
>> umount context.
>
> I _really_ think this is too late, considering how broken it has been.
> We already reverted the WB_SYNC_NONE things exactly because it didn't
> work, didn't we? I'm going to be off-line in two days, and this part
> of the pull request really makes me nervous, if only simply because of
> the history of it all (ie it's always been broken, why shouldn't it be
> broken now?).
>
> IOW, that's a lot of scary changes, that have historically not been
> safe or sufficiently tested, and have caused problems for various
> filesystems. Convince me why they should suddenly be ok to merge?

I agree, it's late and it makes me nervous too. I had them cook for
a day, didn't see any problems. And Christoph would not send it in
unless it passes at least xfs qa, which is what found the problems
last time (the ones we reverted).

It's fixing a regression where umount takes a LONG time if you have
a lot of dirty inodes, since it basically degenerates to a data
integrity writeback instead of a simple WB_SYNC_NONE. If it wasn't
fixing a nasty regression (the distros are all wanting a real fix
for this, it's a user problem), I would not be submitting this code
at this point in time.

--
Jens Axboe

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo(a)vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/
From: Brian Bloniarz on 10 Jun 2010 12:50
On 06/10/2010 12:25 PM, Jens Axboe wrote:
> On 2010-06-10 17:55, Linus Torvalds wrote:
>> On Thu, Jun 10, 2010 at 6:44 AM, Jens Axboe <jaxboe(a)fusionio.com> wrote:
>>>
>>> - A set of patches fixing the WB_SYNC_NONE writeback from Christoph. So
>>> we should finally have both functional and working WB_SYNC_NONE from
>>> umount context.
>>
>> I _really_ think this is too late, considering how broken it has been.
>> We already reverted the WB_SYNC_NONE things exactly because it didn't
>> work, didn't we? I'm going to be off-line in two days, and this part
>> of the pull request really makes me nervous, if only simply because of
>> the history of it all (ie it's always been broken, why shouldn't it be
>> broken now?).
>>
>> IOW, that's a lot of scary changes, that have historically not been
>> safe or sufficiently tested, and have caused problems for various
>> filesystems. Convince me why they should suddenly be ok to merge?
>
> I agree, it's late and it makes me nervous too. I had them cook for
> a day, didn't see any problems. And Christoph would not send it in
> unless it passes at least xfs qa, which is what found the problems
> last time (the ones we reverted).
>
> It's fixing a regression where umount takes a LONG time if you have
> a lot of dirty inodes, since it basically degenerates to a data
> integrity writeback instead of a simple WB_SYNC_NONE. If it wasn't
> fixing a nasty regression (the distros are all wanting a real fix
> for this, it's a user problem), I would not be submitting this code
> at this point in time.
>

Reinforcing that last point: from what I could figure out, Fedora 13
is shipping the buggy WB_SYNC_NONE patch currently. Ubuntu 10.04 is
shipping an in-kernel workaround that has serious performance
drawbacks.

https://bugzilla.kernel.org/show_bug.cgi?id=15906 has links to the
downstream bugs.
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo(a)vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/
From: Christoph Hellwig on 10 Jun 2010 12:50
On Thu, Jun 10, 2010 at 06:25:17PM +0200, Jens Axboe wrote:
> I agree, it's late and it makes me nervous too. I had them cook for
> a day, didn't see any problems. And Christoph would not send it in
> unless it passes at least xfs qa, which is what found the problems
> last time (the ones we reverted).
>
> It's fixing a regression where umount takes a LONG time if you have
> a lot of dirty inodes, since it basically degenerates to a data
> integrity writeback instead of a simple WB_SYNC_NONE. If it wasn't
> fixing a nasty regression (the distros are all wanting a real fix
> for this, it's a user problem), I would not be submitting this code
> at this point in time.

Maybe give it a bit more beating in linux-next and send it off to Linus
once he's back from his vacation?
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo(a)vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/
From: Linus Torvalds on 10 Jun 2010 13:00
On Thu, Jun 10, 2010 at 9:25 AM, Jens Axboe <jaxboe(a)fusionio.com> wrote:
>
> It's fixing a regression where umount takes a LONG time if you have
> a lot of dirty inodes, since it basically degenerates to a data
> integrity writeback instead of a simple WB_SYNC_NONE. If it wasn't
> fixing a nasty regression (the distros are all wanting a real fix
> for this, it's a user problem), I would not be submitting this code
> at this point in time.

I'm not sure if you noticed, we had a separate thread with Dave
Chinner that resulted in three hopefully fairly minimal patches going
in instead.

See commits

git log -3 d87815cb2090

and I thought that last one (first one applied: "pay attention to
wbc->nr_to_write") was the one that had fixed the worst XFS issues.

But maybe it was an unrelated thing.

Linus
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo(a)vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/
 |  Next  |  Last
Pages: 1 2 3 4
Prev: virtio-pci: disable msi at startup
Next: [PATCH 1/1] perf record: Don't call newt functions when not initialized